Business optimisation with artificial intelligence: where to start
«We should do something with AI» — most conversations start with that sentence and end with nothing. The reason is simple: companies try to adopt artificial intelligence in general, rather than in the specific place where it saves hours. Here is the working order we take clients through — built on arithmetic, not enthusiasm.
Start with hours, not with technology
The question «which AI should we adopt» has no answer until another one is answered: what exactly is done by hand in your company every day. So the first step is not choosing a model, it is plain time tracking.
Take one week and write down how many hours the team spends sorting mail, moving data between systems, preparing standard documents, building reports and answering the same customer questions. Multiply by the cost of an hour. You will get a list of processes with a price tag — and you will see that two or three of them cost more than the whole automation project.
The processes that automate most easily
In practice the fastest return comes from boring things — precisely the ones that feel embarrassing to call «innovation»:
- Incoming requests and email. The model reads a message or form, extracts the essentials, fills a CRM record and sets priorities. The manager sees structure instead of a pile of messages.
- Documents. Invoices, delivery notes, contracts: line items, totals and details are extracted into a spreadsheet or accounting system. Human review stays, but it takes minutes.
- Customer support. An assistant trained on your prices and policies covers the standard questions around the clock and hands complex cases to a person.
- Catalogue and content. Product descriptions, categories, filters, supplier price imports.
- Reports. The weekly summary builds itself, with a plain-language explanation of the changes.
- Price monitoring. Watching competitors and suppliers with alerts on deviations.
The payback arithmetic
Our first process turnkey costs from €600, more complex scenarios with several data sources and RAG from €1200, support and model API from €50/month.
The maths is simple: a process eating 20 hours a month at even €10 an hour costs you €2400 a year. A €600 solution with €600 a year of running costs pays for itself in roughly a quarter. If the arithmetic does not work out, the process should not be automated — and an honest contractor says so before the contract, not after.
For comparison: agencies in Germany charge from €5000 for similar work, and large rollouts there run to €20,000–50,000. The technology is the same — the difference is the cost of an hour of the team's time.
Four mistakes that burn the budget
- Starting with the hardest case. «Let us put AI into sales and analytics right away» — and the project drowns in approvals. The right move: one process, the most expensive in hours, with a result in three weeks.
- Automating chaos. If a process is not described, AI will not create order — it will reproduce the mess faster. First a simple written procedure, then automation.
- Not counting before the start. Without an «hours per month» figure there is no way to prove the effect, and the project stays an «interesting experiment».
- Buying a tool instead of a result. A subscription to a fashionable service with no integration into your data dies within two months — people simply stop using it.
Security: what must not leave the building
Sensible adoption starts by splitting the data. Customers' personal data — names, phone numbers, addresses — is masked before it reaches the model. Keys, knowledge bases and logs stay on your side. For EU companies there is a separate setup: EU-hosted models and GDPR-compliant processing.
And the main rule, worth writing into the contract: where a mistake is expensive, AI prepares and a person confirms. That is not distrust of the technology, it is ordinary control — the same you would apply to a new employee.
A plan for the first 30 days
- Week 1. Time tracking and a process audit. The output is a routine map with an «hours per month» figure next to every item.
- Week 2. Pick one process — the most expensive in time and the simplest in data. Describe the desired outcome in a single paragraph.
- Weeks 3–4. Launch on real data with a human at the end. Measure: how long it took before, how long it takes now.
- Day 30. Decide on the numbers: scale, refine or stop. All three answers are legitimate.
According to the Mastercard SME Index, roughly one in five entrepreneurs in Ukraine has already adopted AI. But the winner is not the one who «has AI» — it is the one who removed a specific routine with it and measured the effect.